Teli Ma^1*, Jia Zheng^1*, Zifan Wang¹, Ziyao Gao¹, Jiaming Zhou¹, Junwei Liang^1,2,#

*equal contributions, #corresponding author

¹HKUST(GZ), ²HKUST

[🌐 Project Page] | [📄 GLOVER++ Paper] | [📄 GLOVER Paper] | 🤗 Huggingface Data | [📺 Video] | [🤗 Pretrained Weights]

• GLOVER++ aims to distill actionable affordance knowledge from rich human videos, and demonstrates the effective transfer as an explicit representation for a variety of manipulation tasks.
• We contribute a large-scale affordance-annotated dataset—HOVA-500K, that provides the necessary scale and diversity to learn generalizable affordance representations.
• We present GLOVER++, a global-to-local paradigm of affordance training policy based on HOVA-500K, showing fine-grained affordance representation and generalizable affordance reasoning capability. GLOVER++ achieves state-of-the-art performance in the HOVA-500K evaluation benchmark.
• Extensive applications in tasks like zero-shot manipulation, multi-task imitation learning, long-horizon and bimanual manipulation demonstrate the huge potential of HOVA-500K and GLOVER++.

• We introduce HOVA-500K, a large-scale affordance-annotated dataset constructed from existing human videos and images. The HOVA-500K comprises 500,000 meticulously annotated images spanning 1,726 object categories and 675 action categories, creating a comprehensive taxon- omy of human-object interactions.

• Download the HOVA-500K dataset, Use the following command to merge the dataset splits into a single .tar.gz file:

cat HANDAL/part_* > HANDAL.tar.gz
cat Ego4D/part_* > Ego4D.tar.gz
cat epic-100/part_* > epic-100.tar.gz

• Uncompress these .tar.gz files and organize them as follows:

├── HOVA-500K
│   ├── 3doi
│   │   ├── GT_gaussian
│   │   └── images
│   ├── Ego4D
│   │   ├── GT_gaussian
│   │   └── frames
│   ├── HANDAL
│   │   └── annotations
│   │   |    ├── GT_gaussian_train
│   │   |    └── GT_gaussian_test
|   │   └── images
│   └── epic-100

Note: the "annotations" files should be put in the same directory as the training code.

1. Clone the repository:

git clone https://github.com/TeleeMa/GLOVER.git
cd GLOVER

2. Install dependencies:

We use Python 3.9

pip install -r requirements.txt

3. Download pre-trained models:

• LISA Plus 7B model
• CLIP ViT-L/14 model
• SAM ViT-h
• Place them in the specified directories and configure the model paths in the training script.

Basic training command:

bash train_glover.sh

or Advanced training with GLOVER++:

bash train_glover_plus.sh

NOTE: Key training parameters must be set individually:

• --version: /path/to/LISA_Plus_7b
• --vision-tower: /path/to/clip-vit-large-patch14
• --sam_vit_path: /path/to/sam_vit_h_4b8939.pth (only for GLOVER++)
• --dataset_dir: /path/to/HOVA-500K datasets

When training is finished, to get the full model weight:

cd ./runs/glover(++)/ckpt_model && python zero_to_fp32.py . ../pytorch_model.bin

Merge the LoRA weights of pytorch_model.bin, save the resulting model to your desired path in Hugging Face format:

bash merge_weights.sh

bash eval.sh

NOTE: Key evaluation parameters must be set individually:

• --dataset_dir: /path/to/HOVA-500K datasets
• --version: /path/to/GLOVER(++) model
• --model_arch: Choose from 'glover' or 'glover++'

bash infer.sh

NOTE: Key inference parameters must be set individually:

• --version: Path to GLOVER(++) model
• --model_arch: Choose from 'glover' or 'glover++'
• --image_path: Path to input image
• --objects: Target objects(e.g., 'bottle,cup')
• --actions: Target actions(e.g., 'pick up,raise')

If you find this project useful in your research, please consider citing:

@article{ma2025glover++,
  title={GLOVER++: Unleashing the Potential of Affordance Learning from Human Behaviors for Robotic Manipulation},
  author={Ma, Teli and Zheng, Jia and Wang, Zifan and Gao, Ziyao and Zhou, Jiaming and Liang, Junwei},
  journal={arXiv preprint arXiv:2505.11865},
  year={2025}
}

• We would like to thank the LISA++ and SAM for their contributions.